phoenix/litellm-mirror
0
0
Fork 1
mirror of https://github.com/BerriAI/litellm.git synced 2025-04-27 11:43:54 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Improved O3 + Azure O3 support (#8181)
All checks were successful
Read Version from pyproject.toml / read-version (push) Successful in 13s
Details

Browse source 
* fix: support azure o3 model family for fake streaming workaround (#8162)

* fix: support azure o3 model family for fake streaming workaround

* refactor: rename helper to is_o_series_model for clarity

* update function calling parameters for o3 models (#8178)

* refactor(o1_transformation.py): refactor o1 config to be o series config, expand o series model check to o3

ensures max_tokens is correctly translated for o3

* feat(openai/): refactor o1 files to be 'o_series' files

expands naming to cover o3

* fix(azure/chat/o1_handler.py): azure openai is an instance of openai - was causing resets

* test(test_azure_o_series.py): assert stream faked for azure o3 mini

Resolves https://github.com/BerriAI/litellm/pull/8162

* fix(o1_transformation.py): fix o1 transformation logic to handle explicit o1_series routing

* docs(azure.md): update doc with `o_series/` model name

---------

Co-authored-by: byrongrogan <47910641+byrongrogan@users.noreply.github.com>
Co-authored-by: Low Jian Sheng <15527690+lowjiansheng@users.noreply.github.com>
This commit is contained in:
Krish Dholakia 2025-02-01 09:52:28 -08:00 • committed by GitHub
parent 91ed05df29
commit 23f458d2da
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
14 changed files with 211 additions and 37 deletions
Download patch file Download diff file Expand all files Collapse all files

152
litellm/llms/openai/chat/o1_transformation.py
View file

@ -1,152 +0,0 @@
"""
Support for o1 model family
https://platform.openai.com/docs/guides/reasoning
Translations handled by LiteLLM:
- modalities: image => drop param (if user opts in to dropping param)
- role: system ==> translate to role 'user'
- streaming => faked by LiteLLM
- Tools, response_format => drop param (if user opts in to dropping param)
- Logprobs => drop param (if user opts in to dropping param)
"""
from typing import List, Optional
import litellm
from litellm import verbose_logger
from litellm.litellm_core_utils.get_llm_provider_logic import get_llm_provider
from litellm.types.llms.openai import AllMessageValues, ChatCompletionUserMessage
from litellm.utils import (
supports_function_calling,
supports_response_schema,
supports_system_messages,
)
from .gpt_transformation import OpenAIGPTConfig
class OpenAIO1Config(OpenAIGPTConfig):
"""
Reference: https://platform.openai.com/docs/guides/reasoning
"""
@classmethod
def get_config(cls):
return super().get_config()
def should_fake_stream(
self,
model: Optional[str],
stream: Optional[bool],
custom_llm_provider: Optional[str] = None,
) -> bool:
if stream is not True:
return False
if model is None:
return True
supported_stream_models = ["o1-mini", "o1-preview"]
for supported_model in supported_stream_models:
if supported_model in model:
return False
return True
def get_supported_openai_params(self, model: str) -> list:
"""
Get the supported OpenAI params for the given model
"""
all_openai_params = super().get_supported_openai_params(model=model)
non_supported_params = [
"logprobs",
"top_p",
"presence_penalty",
"frequency_penalty",
"top_logprobs",
]
try:
model, custom_llm_provider, api_base, api_key = get_llm_provider(
model=model
)
except Exception:
verbose_logger.debug(
f"Unable to infer model provider for model={model}, defaulting to openai for o1 supported param check"
)
custom_llm_provider = "openai"
_supports_function_calling = supports_function_calling(
model, custom_llm_provider
)
_supports_response_schema = supports_response_schema(model, custom_llm_provider)
if not _supports_function_calling:
non_supported_params.append("tools")
non_supported_params.append("tool_choice")
non_supported_params.append("parallel_tool_calls")
non_supported_params.append("function_call")
non_supported_params.append("functions")
if not _supports_response_schema:
non_supported_params.append("response_format")
return [
param for param in all_openai_params if param not in non_supported_params
]
def map_openai_params(
self,
non_default_params: dict,
optional_params: dict,
model: str,
drop_params: bool,
):
if "max_tokens" in non_default_params:
optional_params["max_completion_tokens"] = non_default_params.pop(
"max_tokens"
)
if "temperature" in non_default_params:
temperature_value: Optional[float] = non_default_params.pop("temperature")
if temperature_value is not None:
if temperature_value == 1:
optional_params["temperature"] = temperature_value
else:
## UNSUPPORTED TOOL CHOICE VALUE
if litellm.drop_params is True or drop_params is True:
pass
else:
raise litellm.utils.UnsupportedParamsError(
message="O-1 doesn't support temperature={}. To drop unsupported openai params from the call, set `litellm.drop_params = True`".format(
temperature_value
),
status_code=400,
)
return super()._map_openai_params(
non_default_params, optional_params, model, drop_params
)
def is_model_o1_reasoning_model(self, model: str) -> bool:
if model in litellm.open_ai_chat_completion_models and "o1" in model:
return True
return False
def _transform_messages(
self, messages: List[AllMessageValues], model: str
) -> List[AllMessageValues]:
"""
Handles limitations of O-1 model family.
- modalities: image => drop param (if user opts in to dropping param)
- role: system ==> translate to role 'user'
"""
_supports_system_messages = supports_system_messages(model, "openai")
for i, message in enumerate(messages):
if message["role"] == "system" and not _supports_system_messages:
new_message = ChatCompletionUserMessage(
content=message["content"], role="user"
)
messages[i] = new_message # Replace the old message with the new one
return messages